部署人工智能驱动的数字病理学以支持胰腺癌真实世界临床决策

Pancreatic ductal adenocarcinoma (PDAC) remains one of the deadliest malignancies, driven by limited therapeutic options and the absence of widely implemented, clinically actionable biomarkers. Transcriptomic subtyping, particularly the classical (CL) versus basal-like (BL) Moffitt classification, offers prognostic and predictive value: CL tumors show improved outcomes and greater sensitivity to 5 fluorouracil based regimens (e.g., FOLFIRINOX). However, BL tumors exhibit poor outcomes across treatment regimens and may benefit from clinical trial prioritization or intensified oversight. But routine use of RNA-based subtyping is hindered by cost, turnaround time, and restricted access to commercial assays such as Purity Independent Subtyping of Tumors (PurIST). To overcome these barriers, we developed a deep learning model that infers PDAC molecular subtypes directly from hematoxylin and eosin (H&E) whole slide images (WSIs) by integrating with matched RNA-sequencing data for supervised training. In this initial iteration, WSIs from 126 Pancreatic Cancer Action Network (PANCAN) patients with high-quality slides and confirmed CL or BL transcriptomic profiles were curated. Bulk RNA-seq underwent standardized preprocessing, including quality control, alignment, normalization, and quantification of Moffitt-derived gene signatures to generate high-confidence molecular labels for supervised training. These annotations served as ground truth for supervised training. Using five-fold cross-validation, the model classified PDAC tumors into CL and BL subtypes with strong performance (area under curve, AUC: 0.83; accuracy: 77%; specificity: 80%; sensitivity: 72%), comparable to existing image-based PurIST subtyping literature (AUC 0.83-0.86). Our ongoing work with larger multi-institutional datasets aims to further enhance accuracy and generalizability. This proof of concept establishes the feasibility of AI-driven digital pathology for rapid, scalable Moffitt PDAC molecular subtyping directly from WSIs of routine H&E slides. By eliminating the need for RNA-based assays, this approach offers a cost-effective and scalable alternative, particularly valuable for real-world and resource-limited clinical settings. Prospective validation studies will be crucial for refining performance, assessing clinical utility, and enabling integration into precision oncology workflows for PDAC treatment.